A method for predicting highway pavement temperature based on physical constraints

Abstract

The present invention discloses a highway pavement temperature forecasting method based on physical constraints, including the following steps: collecting and preprocessing meteorological data, historical observation data of pavement temperature and geographic data of all stations in the region, adding time information to construct time features; building a U-net neural network model, correcting spatial field data of multi-station multi-meteorological element forecast data, and outputting the corrected meteorological data of each station; building a bidirectional LSTM network based on the U-net model, processing time series data, and using loss functions to evaluate and optimize model performance; adding physical constraints to the model, including the time relationship of temperature changes, seasonal and day and night change characteristics, material thermal properties, thermal balance equations, etc., to optimize model results; using a test set to evaluate the generalization ability and forecast accuracy of the model, and fine-tuning model parameters. This method significantly improves the accuracy and reliability of pavement temperature prediction, and provides a reliable basis for traffic management and safety warning.

Description

Technical Field

[0001] The invention relates to the field of traffic meteorology, and in particular to a method for predicting highway pavement temperature based on physical constraints. Background Art

[0002] Accurate forecasting of highway pavement temperature is crucial for traffic safety. Pavement temperature directly affects the driving performance and safety of vehicles, especially in extreme weather conditions, such as low temperature icing in winter or high temperature tire blowout in summer. Accurate pavement temperature forecasting can help traffic management departments take measures in advance and effectively prevent traffic accidents. However, existing pavement temperature forecasting methods have some limitations and are difficult to meet actual needs. Specifically, they are: Single site data: Most existing pavement temperature forecasting methods are based on meteorological data from a single site only, ignoring the impact of surrounding sites. This method often performs poorly under complex weather conditions because data from a single site cannot fully reflect the meteorological conditions of the entire area. For example, the temperature at one site may be affected by factors such as local terrain and building obstruction, which may not exist at other sites, resulting in deviations in the forecast results.

[0003] Lack of physical constraints: Existing forecasting methods are mostly based on data-driven machine learning algorithms, such as support vector machines (SVM) and random forests (RF). Although these methods can capture the relationship between data to a certain extent, they lack a deep understanding of physical processes. The change in road surface temperature is caused by a series of complex physical processes, including the combined effects of solar radiation, air temperature, wind speed, humidity and other factors. Methods that rely solely on data-driven methods are difficult to accurately reflect these physical processes, resulting in low reliability of forecast results.

[0004] Insufficient spatiotemporal data processing capabilities: Existing methods have limitations in processing spatiotemporal data from multiple sites and multiple meteorological elements. The meteorological conditions along the highway are highly spatially heterogeneous and temporally dynamic, requiring a model that can efficiently process large-scale spatiotemporal data. Traditional statistical methods and simple machine learning models perform poorly in this regard and cannot effectively extract and utilize useful information from these complex data.

[0005] Limited model generalization ability: Existing forecast models often perform well in specific areas or time periods, but have poor generalization ability in other areas or time periods. This is because the dataset used to train the model may not be representative, or the model itself lacks sufficient robustness. In actual applications, the meteorological conditions and road conditions of highways vary with factors such as geographical location and seasonal changes. The model needs to have strong generalization ability to adapt to various situations.

[0006] In the face of the above limitations, there is an urgent need for a pavement temperature prediction method that can comprehensively utilize multi-site data, consider physical processes, efficiently process spatiotemporal data, and have strong generalization capabilities. Summary of the invention

[0007] The purpose of the present invention is to solve the above technical problems and provide a method for predicting highway pavement temperature based on physical constraints.

[0008] In order to achieve the above object, the present invention provides a highway pavement temperature prediction method based on physical constraints, which adopts the following technical solutions:

[0009] A method for predicting highway pavement temperature based on physical constraints comprises the following steps:

[0010] S1. Obtain meteorological data of all stations in the region, obtain grid data of multiple elements divided by longitude and latitude output by numerical weather forecast models in the region; obtain historical observation data and geographic data of road surface temperature;

[0011] S2, preprocessing the acquired original data set, adding time information, and constructing time features;

[0012] S3. Build a U-net neural network model and train it. Based on the U-net model, build a bidirectional LSTM network to process time series data, capture the bidirectional dependencies in the time series, and use the loss function to evaluate and optimize the performance of the model during the training process.

[0013] S4. Add physical constraints to the model, optimize the model results, divide the data into a training set and a test set, and use the training set and the test set to train the model and optimize the parameters; the physical constraints include the time relationship of temperature changes, seasonal and day-night change characteristics, material thermal properties, and thermal balance equations.

[0014] S5. Use the validation set to evaluate the trained model and test the model’s generalization ability and prediction accuracy.

[0015] As a preferred embodiment of the present invention, the specific method of step S2 is:

[0016] Step 21: Clean the collected original data set and process the missing values ​​and outliers;

[0017] Step 22: Add time information and construct time features, which include minutes, hours, days, and months.

[0018] Step 23: Divide the preprocessed data into training set, test set and validation set according to a certain ratio, ready for model training and evaluation.

[0019] As a preferred embodiment of the present invention, the specific method of building a U-net neural network model in step S3 and training the model is:

[0020] The U-net network structure is divided into six parts: input layer, convolution layer, pooling layer, upsampling layer, skip connection layer and output layer;

[0021] The pre-processed multi-site multi-meteorological element forecast data is input into the input layer, convolved through the convolution layer, pooled through the pooling layer, filtered and reduced in dimension to retain the main features of the data;

[0022] The data enters the upsampling layer and is reconstructed layer by layer, and finally the data correction result of the entire spatial field is output.

[0023] As a preferred embodiment of the present invention, in step S3, a bidirectional LSTM network is constructed to process time series data, capture bidirectional dependencies in the time series, and a specific method for using a loss function to evaluate and optimize the performance of the model during training is as follows:

[0024] The road surface temperature at each time point is extracted from the historical observation data, and the corrected meteorological data output from the U-net neural network model, including temperature, radiation, wind speed, cloud cover, humidity, and time information, including minutes, hours, days, and months, are added;

[0025] The input layer receives the prepared data, the bidirectional LSTM layer captures the bidirectional relationship of the time series data, the Dropout layer randomly discards some neurons to reduce the complexity of the model and prevent overfitting, and the output layer generates the future road surface temperature prediction results;

[0026] Use mean squared error as the basic loss function to evaluate the difference between the model's predicted value and the actual value;

[0027]

[0028] in, is the actual road surface temperature, is the road surface temperature predicted by the model, and N is the number of samples;

[0029] Use the training set data to train the model, and optimize the model parameters through the back propagation algorithm to minimize the loss function;

[0030] Use the validation set data to evaluate the performance of the model.

[0031] As a preferred embodiment of the present invention, the method of adding the time relationship constraint of temperature change to the model is:

[0032] An additional penalty term is added to the loss function to constrain the temperature change rate:

[0033] in, is the predicted temperature at time point (t), is a tuning parameter.

[0034] As a preferred embodiment of the present invention, a method for adding seasonal and day-night variation characteristic constraints to the model is: adding seasonal parameters and sunshine duration as additional input features to the model.

[0035] As a preferred embodiment of the present invention, the method of adding material thermal property constraints to the model is: adding thermal conductivity, specific heat capacity and density of the pavement material to the model as additional input features.

[0036] As a preferred embodiment of the present invention, the method of adding the heat balance equation constraint to the model is: adding the penalty term of the heat balance equation to the loss function, and the heat balance equation is expressed as:

[0037] in, is the input heat, is the loss of heat, is the temperature change.

[0038] In summary, the present application includes at least one of the following beneficial technical effects:

[0039] 1. The present invention significantly improves the accuracy and reliability of road surface temperature prediction by constructing a U-net neural network model and a bidirectional LSTM network and adding physical constraints to the model. First, the U-net neural network model can effectively correct the spatial field data of multi-site multi-meteorological element forecast data to a point and output the corrected meteorological data of each site. This process not only improves the accuracy of meteorological data, but also provides high-quality input data for subsequent time series prediction. Secondly, the bidirectional LSTM network can more comprehensively understand the patterns and trends in historical data by capturing the bidirectional dependencies of time series data, thereby improving the accuracy of prediction. Finally, by adding physical constraints such as smoothness constraints and heat balance equations of temperature changes to the loss function, the prediction results of the model are not only more accurate in numerical terms, but also more in line with actual physical laws. This multi-faceted optimization ensures the prediction performance of the model under various meteorological conditions, providing a reliable basis for traffic management and safety warning.

[0040] 2. In the process of model construction, by adding the Dropout layer and physical constraints, this case effectively reduced the risk of overfitting of the model and improved the generalization ability of the model. The Dropout layer reduces the complexity of the model by randomly discarding some neurons and avoids overfitting of the model on the training data. This enables the model to maintain good prediction performance when facing unseen data. In addition, the introduction of physical constraints not only improves the prediction accuracy of the model, but also enhances the robustness of the model. For example, the smoothness constraint of temperature change ensures the smoothness of the prediction results and avoids abrupt temperature changes, which is very important in practical applications. The constraints of the heat balance equation ensure that the prediction results of the model conform to the laws of physics, making the model's predictions under different meteorological conditions and pavement materials more reliable. These measures work together to enable the model to show stronger generalization ability and higher stability in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is a flow chart of an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The following is combined with Figure 1 The present invention is described in further detail.

[0043] Reference Figure 1 This embodiment discloses a method for predicting highway pavement temperature based on physical constraints, comprising the following steps:

[0044] Step S1: Obtain meteorological data of all stations in the region from each traffic meteorological station in the region, including but not limited to temperature, radiation, wind speed, cloudiness, humidity, etc. Obtain grid data of multiple elements divided by longitude and latitude output by numerical weather forecast model in the region; obtain historical observation data of road surface temperature and geographic data, such as surface characteristics and road material, etc.

[0045] Step S2: preprocess the acquired original data set, add time information, and construct time features;

[0046] Step S21, clean the collected original data set and process the missing values ​​and outliers. Data cleaning: Check data integrity: Check whether each field has missing values, and record the number and location of missing values. Process missing values: If a data point has many missing values ​​and is not important, you can choose to delete the data point. For a small number of missing values, you can use interpolation methods (such as linear interpolation, polynomial interpolation), mean / median filling, nearest neighbor filling, etc. to fill them.

[0047] Handling outliers: Use statistical methods (such as Z-score, IQR) or visualization methods (such as box plots) to detect outliers. If the outliers are obviously wrong and there are few of them, you can choose to remove these data points. If the outliers are likely to be measurement errors, you can try to correct these values, such as replacing them with the average of the neighboring values.

[0048] Data consistency check: Ensure data consistency, such as unified timestamp format and consistent units.

[0049] Step S22, add time information and construct time features. Time features include minutes, hours, days, and months. The specific scheme is as follows: Extract timestamp information from the original data to ensure that the timestamp format is unified (such as YYYY-MM-DD HH:MM:SS). Construct time features: extract minute information from timestamps, extract hour information from timestamps, extract date information from timestamps, extract month information from timestamps, etc. Time feature engineering: For periodic features (such as hours, days, and days of the week), sine and cosine transforms can be used to convert them into periodic variables to capture periodic changes. Calculate the time difference between adjacent time points, such as the temperature change between two adjacent hours. Construct sliding window features, such as the average temperature and maximum temperature in the past few hours.

[0050] Step S23: Divide the preprocessed data into a training set, a test set, and a validation set according to a certain ratio, in preparation for model training and evaluation, and determine the division ratio of the data set, for example, 70% (training set), 15% (test set), and 15% (validation set).

[0051] The dataset is divided into training set, test set and validation set using random partitioning method to ensure that the data distribution of each subset is similar.

[0052] Divide them in chronological order, for example, use early data as training set, mid-term data as test set, and late data as validation set.

[0053] Save the divided data sets as separate files or database tables, such as CSV files. Back up the preprocessed data to prevent data loss or damage.

[0054] S3. Build a U-net neural network model and train it. Divide the U-net network structure into six parts: input layer, convolution layer, pooling layer, upsampling layer, skip connection layer and output layer. Input the preprocessed multi-site multi-meteorological element forecast data into the input layer. After several convolutions and pooling, it plays the role of filtering and dimensionality reduction, retaining the main features of the data. After that, the data enters the upsampling process and is reconstructed layer by layer, and finally outputs the data correction result of the entire spatial field.

[0055] Based on the U-net model, a bidirectional LSTM network is constructed to process time series data and capture the bidirectional dependencies in the time series. The loss function is used during the training process to evaluate and optimize the performance of the model.

[0056] The specific steps are:

[0057] Data preparation:

[0058] Historical road surface temperature: The road surface temperature at each time point is extracted from historical observation data. Meteorological data: Corrected meteorological data output from the U-net neural network model, including temperature, radiation, wind speed, cloud cover, humidity, etc. Time features: Add time information, such as minutes, hours, days, months, etc., to help the model better understand time series data.

[0059] Model structure:

[0060] Input layer: receives the prepared data mentioned above.

[0061] Bidirectional LSTM layer: The bidirectional LSTM layer can capture the bidirectional relationship of time series data, that is, it not only considers past data but also future data. This helps the model understand the characteristics of time series more comprehensively.

[0062] Dropout layer: By randomly discarding some neurons, the complexity of the model is reduced to prevent overfitting.

[0063] Output layer: Generates future road surface temperature prediction results.

[0064] Loss function:

[0065] Basic loss function: Use mean squared error (MSE) as the basic loss function to evaluate the difference between the model's predicted value and the actual value.

[0066] in, is the actual road surface temperature, is the road surface temperature predicted by the model, and N is the number of samples;

[0067] Model training:

[0068] Use the training set and test set data to train the model, and optimize the model parameters through the back propagation algorithm to minimize the loss function; use the validation set data to evaluate the performance of the model to ensure that the model can perform well on unseen data.

[0069] Step S4, adding physical constraints to the model, optimizing the model results, dividing the data into a training set and a test set, and using the training set and the test set to train the model and optimize the parameters; the physical constraints include the time relationship of temperature changes, seasonal and day-night change characteristics, material thermal properties, and thermal balance equations.

[0070] The method to add time-dependent constraints on temperature changes in the model is:

[0071] An additional penalty term is added to the loss function to constrain the temperature change rate:

[0072] in, is the predicted temperature at time point (t), is a tuning parameter.

[0073] The method of adding seasonal and day-night variation characteristic constraints to the model is: considering the impact of seasonal variation and day-night temperature difference on temperature, adding seasonal parameters and sunshine duration as additional input features to the model.

[0074] The method of adding material thermal property constraints to the model is as follows: Since this embodiment is based on highway pavement, the pavement material is mostly asphalt, and the thermal conductivity, specific heat capacity and density of asphalt are added as constraints. Changing these properties will affect the temperature change. The thermal conductivity, specific heat capacity and density of the pavement material are added to the model as additional input features.

[0075] The method of adding heat balance equation constraints to the model is: combining factors such as road surface temperature, air temperature, solar radiation, wind speed, humidity, etc., using the heat balance equation to constrain the model, and adding the penalty term of the heat balance equation to the loss function. The heat balance equation is expressed as:

[0076] in, is the input heat, is the loss of heat, is the temperature change.

[0077] Step S5: Use the validation set to evaluate the trained model to test the generalization ability and prediction accuracy of the model.

[0078] The above are all preferred embodiments of the present application, and the protection scope of the present application is not limited thereto. Therefore, any equivalent changes made according to the structure, shape, and principle of the present application should be included in the protection scope of the present application.

Claims

1. A highway pavement temperature prediction method based on physical constraints, characterized in that: The following steps are involved: Step S1, obtaining meteorological data of stations in the region, obtaining grid data of multiple elements divided by longitude and latitude output by numerical weather forecast models in the region; obtaining historical observation data and geographic data of road surface temperature; Step S2: preprocess the acquired original data set, add time information, and construct time features; Step S3, build a U-net neural network model and train the model: divide the U-net network structure into six parts: input layer, convolution layer, pooling layer, upsampling layer, skip connection layer and output layer; The pre-processed multi-site multi-meteorological element forecast data is input into the input layer, convolved through the convolution layer, pooled through the pooling layer, filtered and reduced in dimension to retain the main features of the data; The data enters the upsampling layer, is reconstructed layer by layer, and finally outputs the data correction result of the entire spatial field; Based on the U-net model, a bidirectional LSTM network is constructed to process time series data and capture the bidirectional dependencies in the time series. The loss function is used to evaluate and optimize the performance of the model during the training process: the road surface temperature at each time point is extracted from the historical observation data, and the corrected meteorological data output from the U-net neural network model, including temperature, radiation, wind speed, cloud cover, humidity, and time information, including minutes, hours, days, and months, are added; The input layer receives the above data, the bidirectional LSTM layer captures the bidirectional relationship of the time series data, the Dropout layer reduces the complexity of the model and prevents overfitting by randomly discarding some neurons, and the output layer generates the future road surface temperature prediction results; Use mean squared error as the basic loss function to evaluate the difference between the model's predicted value and the actual value; ； in, is the actual road surface temperature, is the road surface temperature predicted by the model, and N is the number of samples; Use the training set data to train the model, and optimize the model parameters through the back propagation algorithm to minimize the loss function; Use the validation set data to evaluate the performance of the model and adjust the model's hyperparameters based on the evaluation results of the validation set; Step S4, adding physical constraints to the model, optimizing the model results, dividing the data into a training set, a test set, and a validation set, and using the training set and the test set to train and adjust the model; the physical constraints include the time relationship of temperature change, seasonal and day and night change characteristics, material thermal properties, and thermal balance equations; the method of adding the time relationship constraint of temperature change to the model is: An additional penalty term is added to the loss function to constrain the temperature change rate: in, is the predicted temperature at time point (t), is the adjustment parameter; The method of adding seasonal and diurnal variation feature constraints to the model is as follows: adding seasonal parameters and sunshine duration as additional input features to the model; The method of adding material thermal property constraints to the model is as follows: adding the thermal conductivity, specific heat capacity and density of the pavement material as additional input features in the model; The method of adding the heat balance equation constraint to the model is: adding the penalty term of the heat balance equation to the loss function, and the heat balance equation is expressed as: in, is the input heat, is the loss of heat, is the temperature change; Step S5: Use the validation set to evaluate the trained model to test the generalization ability and prediction accuracy of the model.

2. The highway pavement temperature prediction method based on physical constraints according to claim 1 is characterized by: The specific method of step S2 is: Step S21, cleaning the collected original data set and processing the missing values ​​and outliers therein; Step S22: adding time information to construct time features, where the time features include minutes, hours, days, and months; Step S23: Divide the preprocessed data into a training set, a test set, and a validation set according to a certain ratio, ready for model training and evaluation.

Images